Bonsai 27B

Bonsai 27B is an open-source, 27-billion-parameter multimodal model family from PrismML, released July 14, 2026, that ships in extreme low-bit variants — including ternary and 1-bit weights — designed to run natively on phones and laptops rather than in a datacenter. The pitch is that you get a model in the 27B class, historically a server-side weight class, small enough to load and run on an iPhone while keeping performance close to a full-precision version.

It's aimed at developers and researchers building on-device and agentic workflows: local assistants, offline multimodal apps, and privacy-sensitive pipelines where sending data to a hosted API is a non-starter. Because the weights are Apache 2.0, teams can fine-tune, quantize further, and ship Bonsai inside a commercial product without licensing friction — the same freedom that made Llama-class and Mistral open weights popular, but at a footprint that targets edge hardware instead of a GPU box.

What sets Bonsai apart is the bet on ternary and 1-bit quantization as a first-class release rather than an afterthought. Most open models ship at full precision and leave community members to produce aggressive quants that lose quality; PrismML is claiming the low-bit variants are the point, positioning Bonsai as the first major 27B-class model engineered to actually fit and run on a phone. Whether the near-full-precision claim holds up under independent benchmarking is the open question — verify the specific quality-vs-size tradeoff against your own eval before betting a product on it.